According to Hinton's paper Dynamic Routing Between Capsules

python main.py

you can switch the is_multi_mnist to change the training set between Mnist and MultiMnist

batch_size = 64, lr = 0.001, beta1 = 0.9, decay_freq = 2k steps, min_lr = 1e-6, after 10k+ steps

is_shift_augmentation = True

is_shift_agumentation = False

• classification

batch_size = 128, lr = 0.001, beta1 = 0.9, after 30k+ training steps

• segmenting

batch_size = 64, lr = 0.001, beta1 = 0.9, decay_freq = 2k steps, min_lr = 1e-6, after 10k+ steps

Here are some images showing the results of segmenting highly overlapping digits:

• image generation condition on class

I generated some images from random DigitCapsule vectors near the unit sphere :

as we can see, it's not so real, but the pervious images show good results. and notice that Figure 4 in the paper is reconstructed from the neighbourhood of real images' embeding, this may mean that the real images didn't fill all the region around the unit sphere. I think this is due to the autoencoder regularization but not the VAE regularization.

• The accuracy metric on MultiMnist may be different from the paper
• Better results will be got if we run more iterations
• The error rate on test set as showed in figures is only evaluated on one random varing test batch, but not the whole test set.